• Introduction
• Problem Definition & Approach
• Evaluation Method
• Offline Evaluation Pricing Strategy
• Solution in details
• Hotel Pricing
• Alibaba
• Meituan
• Airbnb
• Airline
• E-commence
• Reference

Introduction

Pricing is widely used to maximize revenue in different industries such as airline, e-commence, travel and so on. In this summary, I review 3 works from hotel pricing, 1 from airline pricing, and 1 from e-commence.

Problem Definition & Approach

1. Problem and objective
1. Objective: Find optimal price that maximize revenue
2. Among these works, they are trying to find the optimal price for hotels, ancillaries, and products. There are also some difference among their business
1. For Airbnb, they will recommend price to the host. Since in the end the price will be set by the host, the impact of their strategy is also influenced by the that fact some host will ignore their recommendation. In addition, different from OTA, for a property, most of time, there is only one room.
2. For Alibaba and Meituan, they have full ownership to set price for hotel. So they have more flexibility. In addition, for hotel, there are multiple rooms
3. The work from airline pricing and AWS are similar: there are limit number of products. Which means, compared to hotel industry, there aren’t so much variance among the product, so they tried to personalize the price based on users’ context
2. How do they approach it
1. General ideas
1. Except the airline ancillaries pricing, personalized price is set for room(or listing) level and product level(e-commence)
2. For the airline ancillaries pricing, It's a personalized solution based on user specific feature, such as time of departure, length of stay, group size
2. Approaches
1. Traditional Approach: Direct max-rev pricing strategy
1. By tracking how demand varies with respect to price for a large number of identical products, a demand curve F (P) can be estimated, which determines demand as a function of price P. Then the problem of revenue maximization is to find the price P that yields maximal P×F (P). The key to the success of this approach is to get an accurate estimation of the demand function F (P)
2. Compared to the traditional approach, there is no essential difference in terms of methodology. The main difference is that they optimize at a finer granularity
1. Demand function → booking probability estimation
2. Revenue maximization →apply a set of price, then find the max revenue

Evaluation Method

Offline Evaluation Pricing Strategy

A good set of offline evaluation metrics should be tied to the online business metrics that we are trying to drive. In other words, the improvement on our offline evaluation metrics should yield a lift in our business metrics with high likelihood, so that they can be used to guide our offline strategy hyper-parameter tuning and model development.

Main problem of offline evaluation: we do not really have a ground truth of optimal price.

In the work from Airbnb, a set of offline metrics are developed, which were adopted by Meituan, Alibab & Airline company. Below is the logic and definition of these metrics:

Solution in details

Hotel Pricing

Alibaba

1. Two-stages solution
1. Price Elasticity Prediction Model to predict occupancy → forecast the occupancy on a specific night
1. Define a novel demand function that explicitly models the price elasticity of demand for occupancy prediction
2. Three types of factors affecting the price elasticity in the hotel market
1. competitive factors such as price and quality of competitors
2. temporal factors such as events, seasonality, popularity, etc
3. characteristic factors such as hotel star, location and business district
2. Revenue optimization: apply a set of price, then find the max revenue

Meituan

1. Three-stages solution
1. Get reasonable base rate
1. average value of its historical sales
2. occupancy prediction based on DeepFM and seq2seq
1. model:
1. input: data from previous 28day
2. output: occupancy for the next 7 days
2. Metric: MAPE
3. Features
1. basic features including hotel_id, date info, sale history, average daily rate
2. Statistical feature including booking info and sale info
3. DNN model to model human expertise, and this model will output the recommended price
1. Features
1. base rate
2. predicted occupancy
3. other relating factors
2. data
1. Assumption: We assume that the prices of such samples are adjusted by hotel managers to increase the revenue and the pricing strategy considering a variety of situational factors can be transferred to other hotels.
2. How: we select hotel samples of which the revenues in the period increase compared with their revenues of previous periods, eg., last week, last month or last year.

Airbnb

1. Three components
1. Booking probability estimation → The objective of learning is to predict whether an available future night of a specific listing will be booked as of the day we make this prediction. Here are the example of features
1. listing features
2. temporal features
3. supply and demand dynamics
2. Strategy model (regression model) 1. booking probability is one of the feature
3. Personalization

Airline

1. Approach: 2 components solutions
1. purchase probability model
2. revenue optimization model: given the purchase probability, recommend the optimal price that maximizes the expected revenue
2. Model (compared 3 models)
1. a two-stage
1. purchase probability prediction
2. Since price are within a range due business constrains -> map purchase probability to price
2. a two-stage model
1. uses a deep neural network for forecasting
2. coupled with a revenue maximization technique using discrete exhaustive search (define several percentage beforehand)
3. a single-stage end-to-end deep neural network that recommends the optimal price with customized loss function

E-commence

1. 3 components
1. price elasticity of demand with double ML
2. forecast demand and price
3. Solve the optimization problem